.coderabbit.yaml

# CodeRabbit configuration for Claude Octopus
# Docs: https://docs.coderabbit.ai/reference/configuration

language: en-US

reviews:
  # Block merge until CodeRabbit comments are resolved
  request_changes_workflow: true
  high_level_summary: true
  review_status: true
  commit_status: true
  profile: assertive
  auto_review:
    enabled: true
    drafts: false
    base_branches:
      - main

  path_filters:
    # Skip generated/vendored content
    - "!vendors/**"
    - "!node_modules/**"
    - "!*.lock"

  path_instructions:
    # ── Shell scripts ──────────────────────────────────────────────────
    - path: "scripts/**/*.sh"
      instructions: |
        Shell script conventions for this project:
        - Shebang must be `#!/usr/bin/env bash`
        - Error handling: prefer `set -euo pipefail` (or `set -eo pipefail` for sourced files)
        - CRITICAL: Never use `grep -q` with `set -eo pipefail` — it causes SIGPIPE.
          Use `grep -c ... >/dev/null` instead, or guard with `|| true`.
        - When using `grep -c` as a fallback, use `|| true` not `|| echo "0"` to avoid
          multiline values that break numeric comparisons.
        - Standard dir detection: `SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"`
        - Use `log INFO|WARN|ERROR|DEBUG` for output, not raw echo (except in test scripts).
        - Cleanup via trap: `trap 'rm -rf "$tmpdir"' EXIT INT TERM`

    - path: "hooks/*.sh"
      instructions: |
        Hook scripts run during git operations (pre-push, etc.):
        - Never use `git commit --amend` in hooks — it rewrites history during push.
        - Hooks must exit 0 on success, non-zero to abort the operation.
        - Keep hooks fast — they block the user's git workflow.

    # ── Skill files ────────────────────────────────────────────────────
    - path: ".claude/skills/*.md"
      instructions: |
        Skill file conventions:
        - YAML frontmatter is required with fields: name, description, trigger
        - Description MUST be 120 characters or fewer (enforced by test suite).
        - Description MUST NOT contain: "independent", "compound", "team of teams", "claude instances"
        - Trigger section format: "AUTOMATICALLY ACTIVATE when..." followed by "DO NOT activate for:"
        - When referencing orchestrate.sh, use `${CLAUDE_PLUGIN_ROOT}/scripts/orchestrate.sh`
        - Skills that invoke orchestrate.sh should use the Validation Gate Pattern
          (execution_mode: enforced, pre_execution_contract, validation_gates).

    # ── Persona/agent files ────────────────────────────────────────────
    - path: "agents/personas/*.md"
      instructions: |
        Persona file conventions:
        - Required frontmatter: name, description, model, memory, tools
        - model should be one of: sonnet, inherit, claude-opus-4.6
        - memory should be: user (cross-project) or project (isolated)
        - tools must include at minimum: Read, Glob, Grep
        - Description should mention "Use PROACTIVELY" to guide agent routing.
        - Include when_to_use and avoid_if sections.

    # ── Test files ─────────────────────────────────────────────────────
    - path: "tests/**/*.sh"
      instructions: |
        Test conventions:
        - Source test framework: `source "$SCRIPT_DIR/../helpers/test-framework.sh"`
        - Use test_suite, test_case, test_pass, test_fail assertions.
        - Do not hardcode version numbers in assertions — use >= comparisons.
        - Temp dirs: `TEST_TMP_DIR="/tmp/octopus-tests-$$"` with cleanup trap.
        - grep -q is acceptable in test scripts (they use `set -e` not `set -eo pipefail`).

    # ── orchestrate.sh (main engine) ──────────────────────────────────
    - path: "scripts/orchestrate.sh"
      instructions: |
        This is the main orchestration engine (~19K lines). Key conventions:
        - Cannot be sourced safely (has main execution block at bottom).
        - Feature flags: SUPPORTS_* vars declared ~line 400-460.
        - Model validation: `validate_model_allowed()` uses `grep -Fc` for literal matching.
        - `get_agent_model()` wraps `_get_agent_model_raw()` — model restriction validates here.
        - SIGPIPE-safe: use `grep -c` not `grep -q` throughout.
        - JSON output: escape double quotes in variables before embedding in JSON strings.

    # ── Routing ────────────────────────────────────────────────────────
    - path: "scripts/lib/routing.sh"
      instructions: |
        Routing patterns use compound regex terms to avoid false positives.
        Example: use `marketing.?strategy` not bare `marketing`.
        New patterns should be added to recommend_persona_agent() with
        compound terms that distinguish intent from casual mentions.

    # ── GitHub workflows ───────────────────────────────────────────────
    - path: ".github/workflows/*.yml"
      instructions: |
        CI pipeline runs: smoke → unit → integration → e2e (cascading).
        All scripts must be chmod +x before execution.
        Test artifacts upload as JUnit XML via actions/upload-artifact@v4.

    # ── Marketplace/plugin metadata ────────────────────────────────────
    - path: ".claude-plugin/*.json"
      instructions: |
        Plugin metadata files. marketplace.json is auto-synced by sync-marketplace.sh.
        Version in plugin.json is the source of truth — all other version references
        (README badge, marketplace description) must match.